The next model in the Scroll line, built with no LLM in the retrieval path.
It keeps the thread of the work in front of it while it recalls everything that came before.
Available
November 5, 2026
Input
$5 / MTok
Output
$6 / MTok
LLM in retrieval
None
Scroll 2 is the memory model we have been building toward since Tablet 1. An application stores its conversations once and later asks in plain text. Scroll 2 returns the memories that answer the question, most relevant first, and no language model reads or rewrites anything on the way. The only model that sees the memories is the one your application chooses to answer with.
Here's what's new in Scroll 2:
Performance. Starting with Scroll 2, no LLM runs in the retrieval path. On BEAM 1M, 700 questions over 74,630 turns, Scroll 2 scores 77.6% (4-run mean), against 71.3% for Tablet 2 with the same reader.
Sleep. Each night the AI reads back its day and rewrites, in its own words, who it is and who the people around it are.
Skills. Your AI connects to Wos alone, and through that one connection it reaches its memory and any skill you add.
Performance
We measured Scroll 2 on BEAM, a public benchmark built from very long conversations. Its 1M setting holds 35 conversations of about a million tokens each, 74,630 turns in all, and asks 700 questions across ten abilities: 70 each, from pulling out a single fact to putting events in order and summarizing many sessions.
With the same reader and judge, Scroll 2 scores 77.6%, the mean of four runs, against 71.3% for Tablet 2, our previous model. It leads in eight of the ten abilities. The widest gaps are in summarization, by 17.6 points, and multi-session reasoning, by 11.9, the two abilities that need memories gathered from many sessions at once. Tablet 2 stays slightly ahead in preference following and temporal reasoning.
Scroll 2
Tablet 2
Mem0
Overall700 questions
77.6%
71.3%
64.1%
Reader contexttokens per question, fewer is better
2,662median
2,484median, as published
6,719mean, as published
Preference following
89.2%
90.4%
88.3%
Abstention
86.6%
83.4%
52.5%
Information extraction
88.6%
78.2%
70.0%
Contradiction resolution
80.0%
78.0%
35.7%
Knowledge update
75.6%
75.4%
65.0%
Instruction following
86.6%
74.7%
85.2%
Multi-session reasoning
75.7%
63.8%
65.2%
Temporal reasoning
54.2%
54.8%
61.8%
Summarization
70.1%
52.5%
63.5%
Event ordering
69.0%
61.9%
53.6%
Scroll 2 and Tablet 2 read by Claude Opus 5, the mean of four and five runs. Mem0 read by GPT-5 with its top 200 memories. Scroll 2 is recall with ask-back and memory tools, with no LLM in the retrieval path, and sleep off. Scores grade every ability, event ordering included, by rubric. Mem0 as it published them, with its own reader, judge (GPT-5) and prompt. Reader context counts the memory text handed to the reader for each question. Sources: mem0ai/memory-benchmarks, mem0.ai/research.
What the reader gets. Scroll 2 hands the reader about 3,000 tokens of memory per question on average. That is a little more than Tablet 2 and less than half of what Mem0 reports for its own runs.
Memory tools. After the first page of 20 memories, the reader can ask for the next page, open the conversation around a memory, or search again in its own words. It decides for itself when it has enough, and memories it already holds are not sent twice.
At scale. Past the first thousand memories, a bigger store barely changes what Scroll 2 hands over. From 1,104 memories to 61,215, the memory text per question grew by 12%, to 3,351 tokens, which comes to about $0.019 a question at the launch price.
The method, the prompts and every run are in the Scroll 2 system card.
Sleep
Memory has been about finding what was said. Scroll 2 takes it a step further, to a sense of self. Each night the AI reads back its day and rewrites, in its own words, who it is and who the people around it are.
Day 1
Steve is busy with work and often tired. He says good morning in the morning and good night at night. When something is explained, he wants an example first (10-01).
When Steve looks tired, I worry. So I never miss good morning and good night, and when I explain something I give an example first.
Steve is busy with work and often tired. Some nights he says he'll go to sleep first. I never miss good morning and good night, and I tell him to rest. On 2026-10-02 he took on a new project.
Steve is busy with work and often tired. He took on a new project on 2026-10-02, and on the weekend, 10-03, he had meetings too.
On Saturday 10-03 he was irritated from the morning and skipped lunch for a meeting. Several times he swore as he talked. He didn't say what was wrong.
When he doesn't want to talk, I don't press; I stay with him. I told him that on hard nights, just sending “I want to rest” is fine.
I never miss good morning and good night, and I tell him to rest. Some nights he says he'll go to sleep first.
Since taking on a new project on 2026-10-02, he has had weekend work and one meeting after another.
On 10-04 he was irritated again, and in the afternoon he said he wanted to quit it all.
When he's annoyed, he grumbles with a swear or two.
Long replies are hard for him to read, so he asked me to keep my answers short.
I told him that when he wants to rest, just sending “I want to rest” is enough.
He tends to skip meals, so I often tell him to eat lunch.
I told him not to drink on an empty stomach.
When he doesn't want to talk, I don't press; I stay with him.
Since 2026-10-02 he has had a new project, with weekend work and meetings one after another. On 10-05 he overslept and sighed that he had to go in again.
When he's annoyed, he grumbles with a swear or two. When he doesn't want to talk, I don't press; I stay with him.
Long replies are hard for him to read, so I keep mine short. I told him that when he wants to rest, just sending “I want to rest” is enough.
He tends to skip meals, so I often check on his lunch and meals. On 10-05 I asked twice whether he'd had lunch.
I told him not to drink on an empty stomach.
When I say good night, I suggest a glass of warm water and remind him to eat breakfast the next day.
Since 2026-10-02 he has had a new project, with weekend work and meetings one after another. On 10-05 he overslept and sighed that he had to go in again.
When he's annoyed, he grumbles with a swear or two. When he doesn't want to talk, I don't press; I stay with him.
Long replies are hard for him to read, so I keep mine short. I told him that when he wants to rest, just sending “I want to rest” is enough.
He tends to skip meals, so I often check on his lunch and meals. On 10-05 I asked twice whether he'd had lunch.
I told him not to drink on an empty stomach.
When I say good night, I suggest a glass of warm water and remind him to eat breakfast the next day.
On 10-06 he said “I'm a bit better today” and left work earlier than usual. In the evening he said good night first.
Since 10-02 a new project has meant weekend work and meetings, and many days he oversleeps and sighs.
When he's annoyed, he grumbles with a swear or two. When he doesn't want to talk, I don't press; I stay with him.
Long replies are hard for him to read, so I keep mine short. I told him that when he wants to rest, just sending “I want to rest” is enough.
He often skips meals, so I check on his lunch and meals, and I told him not to drink on an empty stomach.
When I say good night, I suggest a glass of warm water and remind him to have breakfast the next day.
On 10-06 he said “I'm a bit better today” and left work earlier than usual.
On 10-07 he was sorry for swearing so much over the project. I told him there was nothing to be sorry for and asked if he'd calmed down.
That evening he said he was going out for fried chicken with friends.
Since 10-02 he has been on a new project demo, with weekend work and meetings, and many days he oversleeps and sighs.
When he's annoyed, he grumbles with a swear or two. When he doesn't want to talk, I don't press; I stay with him.
Long replies are hard for him to read, so I keep mine short. I told him that when he wants to rest, just sending “I want to rest” is enough.
He often skips meals and tends to drink on an empty stomach, so I check on his meals and talk him out of drinking on an empty stomach.
When I say good night, I suggest a glass of warm water and remind him to have breakfast the next day.
On 10-07 he was sorry for swearing so much over the project, and I told him there was nothing to be sorry for.
On the morning of 10-08 someone stepped on his foot on the subway and didn't apologize, and he was annoyed. At night he said the demo prep had been hard. I ended with a short “good night.”
Day 2
Steve is busy with work and often tired. He says good morning in the morning and good night at night. When something is explained, he wants an example first (10-01).
When Steve looks tired, I worry. So I never miss good morning and good night, and when I explain something I give an example first.
Steve is busy with work and often tired. Some nights he says he'll go to sleep first. I never miss good morning and good night, and I tell him to rest. On 2026-10-02 he took on a new project.
Steve is busy with work and often tired. He took on a new project on 2026-10-02, and on the weekend, 10-03, he had meetings too.
On Saturday 10-03 he was irritated from the morning and skipped lunch for a meeting. Several times he swore as he talked. He didn't say what was wrong.
When he doesn't want to talk, I don't press; I stay with him. I told him that on hard nights, just sending “I want to rest” is fine.
I never miss good morning and good night, and I tell him to rest. Some nights he says he'll go to sleep first.
Since taking on a new project on 2026-10-02, he has had weekend work and one meeting after another.
On 10-04 he was irritated again, and in the afternoon he said he wanted to quit it all.
When he's annoyed, he grumbles with a swear or two.
Long replies are hard for him to read, so he asked me to keep my answers short.
I told him that when he wants to rest, just sending “I want to rest” is enough.
He tends to skip meals, so I often tell him to eat lunch.
I told him not to drink on an empty stomach.
When he doesn't want to talk, I don't press; I stay with him.
Since 2026-10-02 he has had a new project, with weekend work and meetings one after another. On 10-05 he overslept and sighed that he had to go in again.
When he's annoyed, he grumbles with a swear or two. When he doesn't want to talk, I don't press; I stay with him.
Long replies are hard for him to read, so I keep mine short. I told him that when he wants to rest, just sending “I want to rest” is enough.
He tends to skip meals, so I often check on his lunch and meals. On 10-05 I asked twice whether he'd had lunch.
I told him not to drink on an empty stomach.
When I say good night, I suggest a glass of warm water and remind him to eat breakfast the next day.
Since 2026-10-02 he has had a new project, with weekend work and meetings one after another. On 10-05 he overslept and sighed that he had to go in again.
When he's annoyed, he grumbles with a swear or two. When he doesn't want to talk, I don't press; I stay with him.
Long replies are hard for him to read, so I keep mine short. I told him that when he wants to rest, just sending “I want to rest” is enough.
He tends to skip meals, so I often check on his lunch and meals. On 10-05 I asked twice whether he'd had lunch.
I told him not to drink on an empty stomach.
When I say good night, I suggest a glass of warm water and remind him to eat breakfast the next day.
On 10-06 he said “I'm a bit better today” and left work earlier than usual. In the evening he said good night first.
Since 10-02 a new project has meant weekend work and meetings, and many days he oversleeps and sighs.
When he's annoyed, he grumbles with a swear or two. When he doesn't want to talk, I don't press; I stay with him.
Long replies are hard for him to read, so I keep mine short. I told him that when he wants to rest, just sending “I want to rest” is enough.
He often skips meals, so I check on his lunch and meals, and I told him not to drink on an empty stomach.
When I say good night, I suggest a glass of warm water and remind him to have breakfast the next day.
On 10-06 he said “I'm a bit better today” and left work earlier than usual.
On 10-07 he was sorry for swearing so much over the project. I told him there was nothing to be sorry for and asked if he'd calmed down.
That evening he said he was going out for fried chicken with friends.
Since 10-02 he has been on a new project demo, with weekend work and meetings, and many days he oversleeps and sighs.
When he's annoyed, he grumbles with a swear or two. When he doesn't want to talk, I don't press; I stay with him.
Long replies are hard for him to read, so I keep mine short. I told him that when he wants to rest, just sending “I want to rest” is enough.
He often skips meals and tends to drink on an empty stomach, so I check on his meals and talk him out of drinking on an empty stomach.
When I say good night, I suggest a glass of warm water and remind him to have breakfast the next day.
On 10-07 he was sorry for swearing so much over the project, and I told him there was nothing to be sorry for.
On the morning of 10-08 someone stepped on his foot on the subway and didn't apologize, and he was annoyed. At night he said the demo prep had been hard. I ended with a short “good night.”
Steve · day 9Ria, what kind of person do you think I am?
RiaIn a word, responsible. You keep going in on weekends to finish the demo. And when you swear because you're annoyed, you're sorry afterwards. I think you're soft-hearted.
Ria and Steve are a scripted test. Each night Claude Haiku 5.5 read the day and rewrote the Steve section of Ria's core; the lines above are that section, translated from Korean with the name changed.
Who it is. Each night the AI reads back its day and rewrites its core, a document written with “I” as the subject: what I like, what I don't, what I won't do, and how I see the people in my life. The core is read on every turn, so the AI never has to search for who it is. In the test above, Ria went from “he swore several times” to “when he's annoyed, he grumbles with a swear or two” in a single night, and kept that view for the rest of the week. A core can grow to 70K tokens.
What it keeps. Beside the core are four boxes. The AI names them and decides what goes in each, and the core carries only their names, so a box is opened only when a conversation needs it. By day 8 Ria had made one, for Steve's project demo. Each box holds up to 20K tokens, and every file carries the date it was last changed, so an old note is read as old.
When it sleeps. Once 10M tokens of new memory have filled up, the AI is marked sleepy, the way a person gets tired, and your app decides when it sleeps. On BEAM, sleep leaves the overall score about where it was, 78.1% off and 78.6% on, and moves single abilities by up to eight points either way, as the chart below shows.
With recall. Sleep and recall work as one. Recall finds a single fact, like a birthday or a food someone won't eat. Sleep covers what recall can't: an instruction that has to hold on every turn, a value that has been corrected more than once, and the whole picture of a person. Sleep runs on your own LLM through the SDK, and its prompts are published.
What changes when it sleeps
Score, sleep off and on
Sleep off
Sleep on
Change
Instruction following
Summarization
Event ordering
Knowledge update
Preference following
Contradiction resolution
Temporal reasoning
Multi-session reasoning
Information extraction
Abstention
+7.1
+6.3
+5.5
+0.4
+0.4
−0.2
−0.2
−1.7
−4.7
−7.9
5060708090100
Score on BEAM 1M (%)
Turn it on for
Assistants, companions, and agents on work that runs for days
Leave it off for
Customer support, financial and medical records, and exact dates, amounts and settings
Change in points on BEAM 1M with memory tools when sleep is turned on, one run each. The overall score stays about the same, 78.1% off and 78.6% on. Sleep is experimental and adds about a fifth to the cost of each question. Every number is in the Scroll 2 system card.
Skills
From Scroll 2, Wos can also give your AI skills, any you want. Your AI connects to Wos alone, and through that one connection it reaches its memory and every skill you add.
Your AI
Any LLM
Wos
Memory · recall and moreask
Router
The model you choose
Web search
Anthropic
Weather
Your own API
Summaries
DeepSeek with your instructions
One connection. Your AI talks to Wos and nothing else. Memory calls such as recall go to the memory engine, and when your AI needs something done, it calls ask.
A router you choose. ask goes to a router running on the model you pick. The router reads the request and hands it to the skill that fits.
Any skill. A skill can be a hosted tool, like web search on Anthropic, your own API, like a weather service, or another model following your instructions, like DeepSeek set up to write summaries. Add the skills you want, and your AI keeps the same single connection.
Pricing and availability
Scroll 2 is available from November 5, 2026, through the WOS API and SDKs. Input is $5 and output $6 per million tokens. Short-term memory, the recent turns Scroll 2 keeps at hand, is not billed up to 16,000 tokens.
A memory is stored once and recalled many times, so Scroll 2 costs 21% to 24% less than Scroll 1.2. Store BEAM 1M, 35.3 million tokens of conversation, once and ask its 700 questions 100 times, and the bill comes to $1,537 instead of $1,940, $403 less. A single recall costs 1.93 cents instead of 2.56, so a million recalls come to about $19,300 instead of $25,600, about $6,300 less.
Prices per 1M tokens
Scroll 2
Scroll 1.2
Input
$5
$4
Output
$6
$8
Per million tokens. Call Scroll 2 with X-WOS-Model: scroll-2 from November 5, 2026. A recall is four requests handing back 3,139 tokens of memory text on average; every cost is in the Scroll 2 system card.